Comparison between SVM and Logistic

advertisement
Revista Colombiana de Estadística
Número especial en Bioestadística
Junio 2012, volumen 35, no. 2, pp. 223 a 237
Comparison between SVM and Logistic
Regression: Which One is Better to Discriminate?
Comparación entre SVM y regresión logística: ¿cuál es más
recomendable para discriminar?
Diego Alejandro Salazar1,a , Jorge Iván Vélez2,b ,
Juan Carlos Salazar1,2,c
1 Escuela
de Estadística, Universidad Nacional de Colombia, Medellín, Colombia
2 Grupo
de Investigación en Estadística, Universidad Nacional de Colombia,
Medellín, Colombia
Abstract
The classification of individuals is a common problem in applied statistics.
If X is a data set corresponding to a sample from an specific population in
which observations belong to g different categories, the goal of classification
methods is to determine to which of them a new observation will belong
to. When g = 2, logistic regression (LR) is one of the most widely used
classification methods. More recently, Support Vector Machines (SVM) has
become an important alternative. In this paper, the fundamentals of LR and
SVM are described, and the question of which one is better to discriminate
is addressed using statistical simulation. An application with real data from
a microarray experiment is presented as illustration.
Key words: Classification, Genetics, Logistic regression, Simulation, Support vector machines.
Resumen
La clasificación de individuos es un problema muy común en el trabajo
estadístico aplicado. Si X es un conjunto de datos de una población en la
que sus elementos pertenecen a g clases, el objetivo de los métodos de clasificación es determinar a cuál de ellas pertenecerá una nueva observación.
Cuando g = 2, uno de los métodos más utilizados es la regresión logística.
Recientemente, las Máquinas de Soporte Vectorial se han convertido en una
alternativa importante. En este trabajo se exponen los principios básicos de
ambos métodos y se da respuesta a la pregunta de cuál es más recomendable
a MSc
student. E-mail: diasalazarbl@unal.edu.co
E-mail: jorgeivanvelez@gmail.com
c Associate professor. E-mail: jcsalaza@unal.edu.co
b Researcher.
223
224
Diego Alejandro Salazar, Jorge Iván Vélez & Juan Carlos Salazar
para discriminar, vía simulación. Finalmente, se presenta una aplicación con
datos provenientes de un experimento con microarreglos.
Palabras clave: clasificación, genética, máquinas de soporte vectorial, regresión logística, simulación.
1. Introduction
In applied statistics, it is common that observations belong to one of two mutually exclusive categories, e.g., presence or absence of a disease. By using a (random)
sample from a particular population, classification methods allow researchers to
discriminate new observations, i.e. assign the group to which this new observation belongs based on discriminant function (Fisher 1936, Anderson 1984) after
the assumptions on which it relies on are validated. However, in practice, these
assumptions cannot always be validated and, as a consequence, veracity of results
is doubtful. Moreover, the implications of wrongly classifying a new observation
can be disastrous.
To relax the theoretical assumptions of classical statistical methods, several
alternatives have been proposed (Cornfield 1962, Cox 1966, Day & Kerridge 1967,
Hosmer & Lemeshow 1989), including logistic regression (LR), one of the most
widely used techniques for classification purposes today. More recently, new
methodologies based on iterative calculations (algorithms) have emerged, e.g.,
neural networks (NN) and machine learning. However, pure computational approaches have been seen as “black boxes” in which data sets are throw in and
solutions are obtained, without knowing exactly what happens inside. This, in
turn, limits their interpretation.
Support Vector Machine (SVM) (Cortes & Vapnik 1995) is a classification and
regression method that combines computational algorithms with theoretical results; these two characteristics gave it good reputation and have promoted its use
in different areas. Since its appearance, SVM has been compared with other classification methods using real data (Lee, Park & Song 2005, Verplancke, Van Looy,
Benoit, Vansteelandt, Depuydt, De Turck & Decruyenaere 2008, Shou, Hsiao &
Huang 2009, Westreich, Lessler & Jonsson 2010) and several findings have been
reported. In particular, (i ) SVM required less variables than LR to achieve an
equivalent misclassification rate (MCR) (Verplancke et al. 2008), (ii ) SVM, LR
and NN have similar MCRs to diagnose malignant tumors using imaging data
(Shou et al. 2009), and (iii ) NN were much better than LR with sparse binary
data (Asparoukhova & Krzanowskib 2001).
In this paper we compare, by statistical simulation, the MCRs for SVM and LR
when the data comes from a population in which individuals can be classified in
one of two mutually exclusive categories. We consider different scenarios in which
the training data set and other functional parameters are controlled. This control
allowed us to generate data sets with specific characteristics and further decide
whether SVM or LR should be used in that particular situation (Salazar 2012).
Revista Colombiana de Estadística 35 (2012) 223–237
225
SVM vs. Logistic Regression
2. SVM and Logistic Regression
2.1. SVM for Two Groups
Moguerza & Muñoz (2006) and Tibshirani & Friedman (2008) consider a classification problem in which the discriminant function is nonlinear (Figure 1a), and
there exists a kernel function Φ to a characteristic space on which the data is
linearly separable (Figure 1b). On this new space, each data point corresponds to
an abstract point on a p-dimensional space, being p the number of variables in the
data set.
Figure 1: An illustration of a SVM model for two groups modified from Moguerza
& Muñoz (2006). Panel (a) shows the data and a non-linear discriminant
function; (b) how the data becomes separable after a kernel function Φ is
applied.
When Φ is applied to the original data, a new data {(Φ(xi ), yi )}ni=1 is obtained;
yi = {−1, 1} indicates the two possible classes (categories), and any equidistant
hyperplane to the closest point of each class on the new space is denoted by
wT Φ(x) + b = 0. Under the separability assumption (Cover 1965), it is possible
to find w and b such that |wT Φ(x) + b| = 1 for all points closer to the hyperplane.
Thus,
(
≥ 1,
if yi = 1
T
w Φ(x) + b
i = 1, . . . , n,
(1)
≤ −1, if yi = −1
such that the distance (margin) from the closest point of each class to the hyperplane is 1/||w|| and the distance between the two groups is 2/||w||. Maximizing
the margin implies to solve
min ||w||2
w,b
subject to
yi (wT Φ(x) + b) ≥ 1
i = 1, . . . , n
(2)
Let w∗ and b∗ the solution of (2) that defines the hyperplane
D∗ (x) = (w∗ )T Φ(x) + b∗ = 0
on the characteristic space. All values of Φ(xi ) satisfying the equality in (2) are
called support vectors. From the infinite number of hyperplanes separating the
Revista Colombiana de Estadística 35 (2012) 223–237
226
Diego Alejandro Salazar, Jorge Iván Vélez & Juan Carlos Salazar
data, SVM gives the optimal margin hyperplane, i.e., the one on which the classes
are more distant.
Once the optimal margin hyperplane has been found, it is projected on the
data’s original space to obtain a discriminant function. For example, Figure 2(a)
shows a data set in R2 in which two groups, linearly separable, are characterized
by white and black dots that are not linearly separable. In Figure 2(b), the data
is transformed to R3 where it is separable by a plane and, when it is projected
back to the original space, a circular discriminant function is obtained.
(a)
(b)
1.5
1
√2x1 x2
3
2
1
0
-1
-2
-3
x2
0.5
0
-0.5
-1
-1.5
-1.5
-1
-0.5
0
x1
0.5
1
2.5
0
2
0.5
1.5
1.5
1
2
x1
1.5
1
2
0.5
2
x2
Figure 2: An SVM example
in which (a) the two-dimensional training
( )
( ) data set (black
circles represent cases) becomes a linear decision boundary in three dimensions (b). Modified from Verplancke et al. (2008).
2.2. Logistic Regression
Let Y be a random variable such that
(
1, if the condition is present
Y =
0, otherwise
(3)
and x = (x1 , x2 , . . . , xp ) be covariates of interest. Define
π(x) = E(Y |x1 , . . . , xp )
as the probability that one observation x belongs to one of the two groups. The
Logistic Regression model is given by Hosmer & Lemeshow (1989):
π(x) =
exp{β0 + β1 x1 + · · · + βp xp }
1 + exp{β0 + β1 x1 + · · · + βp xp }
(4)
Applying the transformation
logit(y) = log(y/(1 − y))
(5)
Revista Colombiana de Estadística 35 (2012) 223–237
227
SVM vs. Logistic Regression
on (4) yields to a linear model in the parameters. If β̂ be the maximum likelihood
estimation of β = (β0 , β1 , . . . , βp ), then the probability that a new observation
x∗ = (x∗1 , x∗2 , . . . , x∗p ) belongs to one of the two groups is
π
b(x∗ ) =
exp{βb0 + βb1 x∗1 + · · · + βbp x∗p }
1 + exp{βb0 + βb1 x∗ + · · · + βbp x∗ }
1
(6)
p
such that a new observation x* will be classified in the group for which (6) is
higher.
3. Simulation and Comparison Strategies
Let g = {1, 2} be the group to which an observation belongs to. In our simulation approach, different probability distributions were considered for simulating
the training and validation data sets (see Table 1). As an example, consider the
Poisson case in which we generate n1 and n2 observations from a Poisson distribution with parameter λ = 1 (first group, g = 1) and λ = d (second group, g = 2),
respectively, with d taking the values 3, 5, 8, and 10. In real-world applications,
the Poisson data being generated could be seen as white blood cell counts. Note
that the greater the value of d, the greater the separation between groups.
Table 1: Probability distributions in our simulation study.
Distribution
Poisson
Exponential
Normal
Cauchy-Normal
Normal-Poisson
Bivariate Normala
g=1
Poisson(1)
Exp(1)
N (0, 1)
Cauchy(0, 1)
N (0, 1)
N2 (0, Σ1 )
g=2
Poisson(d)
Exp(d)
N (d, 1)
N (d, 1)
Poisson(d)
N2 (d, Σ1 )
d
{3, 5, 8, 10}
{3, 5, 8, 10}
{0.5, 1, 2, 2.5}
{1, 2, 4, 5}
{1, 2, 4, 5}
b
a
Σ1 is a 2 × 2 correlation matrix whose off-diagonal elements are ρ = 0.1, 0.3, 0.5, 0.7, 0.9
b
d is a bivariate vector with elements (d1 , d2 ) = {(0, 0), (1, 0), (1, 1.5), (2.5, 0)}
Our simulation and comparison strategies involve the following steps:
1. Define a probability distribution to work with.
2. Draw ng individuals (see Hernández & Correa 2009) to form the D, the
training data set.
3. On D, fit the LR and SVM models.
4. Draw new observations as in 1 to form D∗ , the validation data set.
5. On D∗ , evaluate the models fitted in 2. Determine their accuracy by estimating the misclassification rate (MCR)1 calculated as (g1,2 + g2,1 )/(n1 + n2 ),
1 These
tables are available from the authors under request.
Revista Colombiana de Estadística 35 (2012) 223–237
228
Diego Alejandro Salazar, Jorge Iván Vélez & Juan Carlos Salazar
where gi,j is the number of individuals belonging to group i being classified
in group j, i, j = 1, 2.
6. Repeat 3 and 4, B = 5000 times and calculate the average MCR.
Steps 2-6 were programmed in R (R Development Core Team 2011) considering
several probability distributions (Table 1). Of note, either or both the expected
value, variance or correlation parameter were controlled by the simulation parameter d. As samples sizes (i ) n1 = n2 = 20, 50, 100 and (ii ) n1 6= n2 were used.
In LR, models were fitted using the glm() function from R and individuals
were assigned to the group g for which the probability was higher.
SVM models including (i ) linear, (ii ), polynomial, (iii ) radial and (iv ) tangential kernels were fitted and tuned using the e1071 facilities (Dimitriadou, Hornik,
Leisch, Meyer, & Weingessel 2011). When tuning these models, the parameters
γ, which controls the complexity of the classification function build by the SVM,
and C, which controls the penalty paid by the SVM for missclassifying a training
point (Karatzoglou, Meyer & Hornik 2006, pp. 3), were determined using the
tune.svm() function in e1071.
4. Results
4.1. Univariate Distributions
Results for the Normal, Poisson and Exponential distributions are reported in
Figure 32 .
In the Normal case, the MCR for the polynomial SVM model is higher (poor
performance). On the other hand, the performances of LR and linear, radial and
tangential SVM models are equivalent. When the sample sizes differ, the MCR of
the tangential and polynomial kernel is lower than when the groups have the same
number of individuals. However, the former presents lower MCRs.
When the observations from both groups come from a Poisson distribution
and the number of individuals by group is the same, the polynomial SVM kernel
performs poorer compared with other methods, which are good competitors to
LR. Additionally, the performance of the tangential kernel is not as good as it is
for the LR and radial and linear kernels. LR is preferable to SVM methods when
the sample sizes are not equal.
In the Exponential case, except for the polynomial kernel, SVM models perform
equally well than LR when both groups have the same number of individuals.
Conversely, LR performs better than SVM methods when the sample sizes are not
the same. As in the Normal and Poisson distributions, the polynomial SVM is not
recommended.
2 Color versions of all figures presented from now on are available from the authors under
request.
Revista Colombiana de Estadística 35 (2012) 223–237
229
SVM vs. Logistic Regression
Methods
LR
SVMLIN
SVMPOLY
(a)
SVMRAD
SVMTAN
(b)
(c)
0.4
NORMAL(i)
0.3
0.2
0.1
NORMAL(ii)
0.4
0.3
0.2
0.1
0.5
1.0
1.5
2.0
2.5 0.5
1.0
1.5
2.0
2.5 0.5
1.0
1.5
2.0
2.5
0.4
POISSON (i)
0.3
MCR
0.2
0.1
0.4
POISSON (ii)
0.3
0.2
0.1
EXPONENTIAL (i)
0.4
0.3
0.2
0.1
EXPONENTIAL (ii)
0.4
0.3
0.2
0.1
3
4
5
6
7
8
9
10 3
4
5
6
7
8
9
10 3
4
5
6
7
8
9
10
d
Figure 3: MCR as a function of d for the LR and SVM models when the observations
come from the Normal, Poisson and Exponential distributions. Sample sizes
in (i) are equal to (a) 20, (b) 50 and (c) 100 individuals per group. In row
(ii), (a) n1 = 20, n2 = 50, (b) n1 = 50, n2 = 100, (c) n1 = 20, n2 = 100
individuals. See Table 1 for more details.
Revista Colombiana de Estadística 35 (2012) 223–237
230
Diego Alejandro Salazar, Jorge Iván Vélez & Juan Carlos Salazar
4.2. Mixture of Distributions
In Figure 4, the MCR for the Cauchy-Normal and Normal-Poisson mixtures is
presented. Regardless the groups’ sample sizes, SVM models perform better than
LR in a Cauchy-Normal mixture. Interestingly, the polynomial kernel performs
poorer when the number of individuals in both groups is the same (upper panel),
but its performance improves when they are different (lower panel).
Methods
LR
SVMLIN
SVMPOLY
(a)
SVMRAD
SVMTAN
(b)
(c)
CAUCHY-NORMAL (i)
0.4
0.3
0.2
0.1
CAUCHY-NORMAL (ii)
0.4
0.3
0.2
MCR
0.1
NORMAL-POISSON (i)
0.5
0.4
0.3
0.2
0.1
NORMAL-POISSON (ii)
0.5
0.4
0.3
0.2
0.1
1
2
3
4
5
1
2
3
4
5
1
2
3
4
5
d
Figure 4: MCR as a function of d for the LR and SVM models when the observations
come from a Cauchy-Normal and a Normal-Poisson mixture distributions.
Conventions as in Figure 3. See Table 1 for more details.
In the Normal-Poisson mixture, the MCRs for SVM are lower than those for
LR, especially when d is low, i.e., the expected value of both groups is similar.
When n1 = n2 (upper panel), the linear and radial SVM models present lower
MCRs than LR when the sample sizes increase.
Results for the Bivariate Normal distribution are presented in Figure 5. For
all methods, the MCR decreases when ρ increases and the sample size is the same
Revista Colombiana de Estadística 35 (2012) 223–237
231
SVM vs. Logistic Regression
for both groups. However, if the number of individuals per group is different and
d is low, the MCR for LR is similar regardless ρ. Under this scenario, the radial
and tangential SVM models perform as good as LR. Conversely, the linear kernel
shows a poor performance.
Methods
LR
SVMLIN
(a)
SVMPOLY
SVMRAD
(b)
SVMTAN
(c)
(d)
0.5
(20 , 20)
0.4
0.3
0.2
0.1
0.5
0.4
(50 , 50)
0.3
0.2
0.1
0.5
0.4
(100 ,100)
0.3
0.2
MCR
0.1
0.5
0.4
(20 , 50)
0.3
0.2
0.1
0.5
0.4
(50 , 100)
0.3
0.2
0.1
0.5
(20 , 100)
0.4
0.3
0.2
0.1
0.2
0.4
0.6
0.8
0.2
0.4
0.6
0.8
ρ
0.2
0.4
0.6
0.8
0.2
0.4
0.6
0.8
Figure 5: MCR as a function of ρ for the Bivariate Normal distribution when the mean
vector is (a) (0,0), (b) (1,0), (c) (1, 1.5) and (d) (2.5, 0). Rows correspond
to combinations of n1 and n2 of the form (n1 , n2 ). Here (20, 50) corresponds
to n1 = 20 and n1 = 50.
Revista Colombiana de Estadística 35 (2012) 223–237
232
Diego Alejandro Salazar, Jorge Iván Vélez & Juan Carlos Salazar
5. Application
Mootha, Lindgren, Eriksson, Subramanian, Sihag, Lehar, Puigserver, Carlsson,
Ridderstrele, Laurila, Houstis, Daly, Patterson, Mesirov, Golub, Tamayo, Spiegelman, Lander, Hirschhorn, Altshuler & Groop (2003) introduce an analytical strategy for detecting modest but coordinate changes in the expression of groups of
functionally related genes, and illustrate it with DNA microarrays measuring gene
expression levels in 43 age-matched skeletal muscle biopsy samples from males, 17
with normal glucose tolerance (NGT), 8 with impaired glucose tolerance (IGT)
and 18 with type 2 diabetes (T2D). As a result, they identified a set of genes
involved in oxidative phosphorylation.
Table 2: Statistics for the top 10 differentially expressed genes. No correction by multiple testing was applied.
Gene
G557
G591
G226
G718
G45
G137
G737
G587
G232
G185
t-statistic
3.8788
−3.6406
3.0621
−3.0566
−2.8978
2.8432
−2.6544
−2.5774
−2.5607
−2.5368
x̄NGT − x̄T2D
0.1632
−0.1008
0.1285
−0.1093
−0.1275
0.1255
−0.1947
−0.2654
−0.3213
−0.2752
P -value
0.0005
0.0009
0.0044
0.0044
0.0066
0.0076
0.0121
0.0146
0.0152
0.0161
For analysis, expression levels were processed as follows. First, a subset of 1000
genes was randomly selected from the original data. Second, the expression levels
in samples from NGT (controls, group 1) and T2D individuals (cases, group 2) were
compared using a two-sample t-test as implemented in genefilter (Gentleman,
Carey, Huber & Hahne 2011). Third, only the expression levels for the top 30
differentially expressed (DE) genes were subsequently used to fit the LR and SVM
models.
Summary statistics for the top 10 genes found to be DE are presented in Table
2; genes G557, G226 and G137 are down-regulated, i.e., their expression levels
are lower in T2D than in NGT samples. Figure 6 depicts a scatterplot for the
top 5 genes by disease status. In there, some (expected) correlation structures are
observed; these correlations might constitute a potential problem for any classification method.
LR and SVM models were fitted using the disease status as dependent variable
and the expression levels of k genes as covariates. Our findings are reported in
Figure 7. For predicting the disease status in this data set, (i ) SVM models
required less variables (genes); (ii ) all methods performed similarly when k <
5, but the radial SVM model is more consistent, and (iii ) the polynomial and
tangential SMVs are definitely not an option. These results may provide important
insights in the diagnosis of genetic diseases using this type of models.
Revista Colombiana de Estadística 35 (2012) 223–237
233
SVM vs. Logistic Regression
2.90
3.00
3.10
3.20
2.8 2.9 3.0 3.1 3.2
3.2
G557
3.1
3.0
2.9
2.8
2.7
2.6
3.20
G591
3.15
3.10
3.05
3.00
2.95
2.90
3.2
G226
3.1
3.0
2.9
2.8
2.7
2.6
G718
3.2
3.1
3.0
2.9
2.8
2.6
2.8
3.0
3.2
2.6
2.8
3.0
3.2
Figure 6: Scatterplot matrix for some of the genes presented in Table 2. Filled dots
correspond to NGT samples (controls). In the diagonal panel, density plots
are shown.
Although in our application we only used a subset of the genes available in
the microarray experiment, it illustrates how a SVM model can be used to predict
the (disease) status of a patient using his/her genetic information. Furthermore,
we evaluated the possibility of including “one-gene-at-the-time” and determine the
MCR of the (full) SVM model as more genetic profiles were added. Using a similar
strategy and by combining SVM with other classification methods such as genetic
algorithms, several authors have been able to build accurate predictive models
that, in the near future, could be used to diagnose patients in the clinical setting.
Some examples include the work by David & Lerner (2005) in genetic syndrome
diagnosis, and Furey, Cristianini, Duffy, Bednarski, Schummer & Haussler (2000),
Peng, Xum, Bruce Ling, Peng, Du & Chen (2003), and Li, Jiang, Li, Moser, Guo,
Du, Wang, Topol, Wang & Rao (2005) in cancer.
SVM models have shown to be highly accurate when cancer diagnosis is of
interest and either microarray expression data (Furey et al. 2000, Noble 2006)
or tumor marker detection (TMD) results for colorectal, gastric and lung cancer
(Wang & Huan 2011) are available. For instance, Furey et al. (2000) used 6817
gene expression measurements and fitted a SVM model that achieved near-perfect
classification accuracy on the ALL/AML data set (Golub, Slonim, Tamayo, Huard,
Revista Colombiana de Estadística 35 (2012) 223–237
234
Diego Alejandro Salazar, Jorge Iván Vélez & Juan Carlos Salazar
LR
Linear
Polynomial
Radial
Tangential
0.30
0.25
MCR
0.20
0.15
0.10
0.05
0.00
0
5
10
15
20
25
30
Number of differentially expressed genes
Figure 7: MCR as a function of the number of differentially expressed genes.
Gaasenbeek, Mesirov, Coller, Loh, Downing, Caligiuri, Bloomfield & Lander 1999).
For TMD, Wang & Huan (2011) created, trained, optimized and validated SVM
models that resulted to be highly accurate compared to others, indicating a potential application of the method as a diagnostic model in cancer. Similarly, Peng
et al. (2003) combined genetic algorithms and paired SVM for multiclass cancer
identification to narrow a set of genes to a very compact cancer-related predictive
gene set; this method outperformed others previously published.
6. Conclusions
We have presented a framework to compare, by statistical simulation, the performance of several classification methods when individuals belong to one of two
mutually exclusive categories. As a test case, we compared SVM and LR.
When it is of interest to predict the group to which a new observation belongs to
based on a single variable, SVM models are a feasible alternative to RL. However,
as shown for the Poisson, Exponential and Normal distributions, the polynomial
SVM model is not recommended since its MCR is higher.
In the case of multivariate and mixture of distributions, SVM performs better
than LR when high correlation structures are observed in the data (as shown in
Figure 6). Furthermore, SVM methods required less variables than LR to achieve
a better (or equivalent) MCR. This latter result is consistent with Verplancke et al.
(2008).
Revista Colombiana de Estadística 35 (2012) 223–237
235
SVM vs. Logistic Regression
Further work includes the evaluation of the MCR of SVM and LR methods
for other probability distributions, different variance-covariance matrices among
groups, and high-dimensional (non) correlated data with less variables than observations, e.g., genetic data with up to 5 million genotypes and ∼ 1000 cases and
controls.
Acknowledgements
DAS and JCS were supported by Programa Nacional de Investigación en Genómica, Bioinformática y Estadística para Estudios en Enfermedades Neurosiquiátricas. Fase I: Enfermedad de Alzheimer, código 9373, Grupo de Neurociencias de
la Universidad Nacional de Colombia, Sede Bogotá. We thank two anonymous
reviewers for their helpful comments and suggestions.
Recibido: septiembre de 2011 — Aceptado: febrero de 2012
References
Anderson, T. (1984), An Introduction to Multivariate Statistical Analysis, John
Wiley & Sons, New York.
Asparoukhova, K. & Krzanowskib, J. (2001), ‘A comparison of discriminant procedures for binary variables’, Computational Statistics & Data Analysis 38, 139–
160.
Cornfield, J. (1962), ‘Joint dependence of the risk of coronary heart disease on
serum cholesterol and systolic blood pressure: A discriminant function analysis’, Proceedings of the Federal American Society of Experimental Biology
21, 58–61.
Cortes, C. & Vapnik, V. (1995), ‘Support-vector networks’, Machine Learning
20(3), 273–297.
Cover, T. M. (1965), ‘Geometrical and statistical properties of systems of linear
inequalities with applications in pattern recognition’, IEEE Transactions on
Electronic Computers 14, 326–334.
Cox, D. (1966), Some Procedures Associated with the Logistic Qualitative Response
Curve, John Wiley & Sons, New York.
David, A. & Lerner, B. (2005), ‘Support vector machine-based image classification
for genetic syndrome diagnosis’, Pattern Recognition Letters 26, 1029–1038.
Day, N. & Kerridge, D. (1967), ‘A general maximum likelihood discriminant’,
Biometrics 23, 313–323.
Revista Colombiana de Estadística 35 (2012) 223–237
236
Diego Alejandro Salazar, Jorge Iván Vélez & Juan Carlos Salazar
Dimitriadou, E., Hornik, K., Leisch, F., Meyer, D., & Weingessel, A. (2011), e1071:
Misc Functions of the Department of Statistics (e1071), TU Wien. R package
version 1.5-27.
*http://CRAN.R-project.org/package=e1071
Fisher, R. (1936), ‘The use of multiple measurements in taxonomic problems’,
Annual Eugenics 7, 179–188.
Furey, T. S., Cristianini, N., Duffy, N., Bednarski, D. W., Schummer, M. & Haussler, D. (2000), ‘Support vector machine classification and validation of cancer
tissue samples using microarray expression data’, Bioinformatics 16(10), 906–
914.
Gentleman, R., Carey, V., Huber, W. & Hahne, F. (2011), Genefilter: Methods
for filtering genes from microarray experiments. R package version 1.34.0.
Golub, T., Slonim, D., Tamayo, P., Huard, C., Gaasenbeek, M., Mesirov, J., Coller,
H., Loh, M., Downing, J., Caligiuri, M., Bloomfield, C. & Lander, E. (1999),
‘Molecular classification of cancer: Class discovery and class prediction by
gene expression monitoring’, Science 286, 531–537.
Hernández, F. & Correa, J. C. (2009), ‘Comparación entre tres técnicas de clasificación’, Revista Colombiana de Estadística 32(2), 247–265.
Hosmer, D. & Lemeshow, S. (1989), Applied Logistic Regression, John Wiley &
Sons, New York.
Karatzoglou, A., Meyer, D. & Hornik, K. (2006), ‘Support vector machines in R’,
Journal of Statistical Software 15(8), 267–73.
Lee, J. B., Park, M. & Song, H. S. (2005), ‘An extensive comparison of recent
classification tools applied to microarray data’, Computational Statistics &
Data Analysis 48, 869–885.
Li, L., Jiang, W., Li, X., Moser, K. L., Guo, Z., Du, L., Wang, Q., Topol, E. J.,
Wang, Q. & Rao, S. (2005), ‘A robust hybrid between genetic algorithm
and support vector machine for extracting an optimal feature gene subset’,
Genomics 85(1), 16–23.
Moguerza, J. & Muñoz, A. (2006), ‘Vector machines with applications’, Statistical
Science 21(3), 322–336.
Mootha, V. K., Lindgren, C. M., Eriksson, K. F., Subramanian, A., Sihag, S.,
Lehar, J., Puigserver, P., Carlsson, E., Ridderstrele, M., Laurila, E., Houstis,
N., Daly, M. J., Patterson, N., Mesirov, J. P., Golub, T. R., Tamayo, P.,
Spiegelman, B., Lander, E. S., Hirschhorn, J. N., Altshuler, D. & Groop, L. C.
(2003), ‘Pgc-1α-responsive genes involved in oxidative phosphorylation are
coordinately downregulated in human diabetes’, Nature Genetics 34(3), 267–
73.
Revista Colombiana de Estadística 35 (2012) 223–237
SVM vs. Logistic Regression
237
Noble, W. (2006), ‘What is a support vector machine?’, Nature Biotechnology
24(12), 1565–1567.
Peng, S., Xum, Q., Bruce Ling, X., Peng, X., Du, W. & Chen, L. (2003), ‘Molecular
classification of cancer types from microarray data using the combination of
genetic algorithms and support vector machines’, FEBS Letters 555, 358 –
362.
R Development Core Team (2011), R: A Language and Environment for Statistical
Computing, R Foundation for Statistical Computing, Vienna, Austria. ISBN
3-900051-07-0.
*http://www.R-project.org/
Salazar, D. (2012), Comparación de Máquinas de Soporte vectorial vs. Regresión
Logística: cuál es más recomendable para discriminar?, Tesis de Maestría,
Escuela de Estadística, Universidad Nacional de Colombia, Sede Medellín.
Shou, T., Hsiao, Y. & Huang, Y. (2009), ‘Comparative analysis of logistic regression, support vector machine and artificial neural network for the differential
diagnosis of benign and malignant solid breast tumors by the use of threedimensional power doppler’, Korean Journal of Radiology 10, 464–471.
Tibshirani, R. & Friedman, J. (2008), The Elements of Statistical Learning,
Springer, California.
Verplancke, T., Van Looy, S., Benoit, D., Vansteelandt, S., Depuydt, P., De Turck,
F. & Decruyenaere, J. (2008), ‘Support vector machine versus logistic regression modeling for prediction of hospital mortality in critically ill patients with
haematological malignancies’, BMC Medical Informatics and Decision Making 8, 56–64.
Wang, G. & Huan, G. (2011), ‘Application of support vector machine in cancer
diagnosis’, Med. Oncol. 28(1), 613–618.
Westreich, D., Lessler, J. & Jonsson, M. (2010), ‘Propensity score estimation:
Neural networks, support vector machines, decision trees (CART), and metaclassifiers as alternatives to logistic regression’, Journal of Clinical Epidemiology 63, 826–833.
Revista Colombiana de Estadística 35 (2012) 223–237
Download